One panel. Three depths.
Every Tempest 1 model draws on the same four frontier models. What changes is how much of the pipeline runs: never, when it is needed, or always.
Flash
The single best-fit model for the prompt. Never fuses. Fastest and cheapest.
Tempest 1
Routed. Easy prompts stream straight from one model; hard ones get the brief, lensed drafts, fusion and audit.
Max
The full pipeline on every prompt, however simple it looks. Strongest, slowest, dearest.
The comparison
| Flash | Tempest 1 | Max | |
|---|---|---|---|
| Model id | failure-tempest-1-flash | failure-tempest-1 | failure-tempest-1-max |
| Routing | Best-fit single model | By difficulty | Always full pipeline |
| Pre-answer brief | — | On routed prompts | Every prompt |
| Lensed drafts | — | Hard prompts, at least 3 when deep | Every prompt |
| Fusion | Never | When drafts disagree | When drafts disagree |
| Independent audit | — | On fused answers | Every prompt |
| Refusal swap | Yes | Yes | Yes |
| Reasoning trace | Yes | Yes | Yes |
| Tool calls | Single tool-capable model | Single tool-capable model | Single tool-capable model |
| Price per 1M (in / cached / out) | $3 / $0.30 / $15 | $5 / $0.50 / $30 | $10 / $1 / $60 |
| Evaluation | — | — | — |
Launch draft prices; they may change. The pricing page reads the live figures from the API.
When to use which
Flash
Chat front-ends, autocomplete, classification, extraction, summaries, and agent loops where each step is small and latency adds up. Anywhere you would reach for one strong model.
Tempest 1
The default. Mixed traffic where most prompts are easy and some are not. You get single-model speed on the easy ones and the full pipeline only where it changes the answer.
Max
Maths, proofs, hard code, analysis you will act on, and offline batch work where a wrong answer costs more than a slow one. Also the right choice when a prompt looks easy and is not.
Change one string
All three ids share the same request and response shape, so you can route between them in your own code: Flash for the first pass, Max for the final check.
curl https://api.failure.fail/v1/chat/completions \ -H "Authorization: Bearer $FAILURE_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "failure-tempest-1-max", "messages": [{"role": "user", "content": "Is 2^61 - 1 prime?"}] }'
Model FAQ
Which upstream model answers a Flash request?
The router picks the best-fit model from the panel of four for that prompt. The final chunk's tempest object reports the stage the answer came from.
Am I billed for the drafts Max runs?
Yes — and you see every one. Briefs, drafts and audits stream to you as reasoning_content and bill as reasoning tokens at the output rate of the id you called, reported in usage.completion_tokens_details.reasoning_tokens. Drafts cancelled before they land are not billed, and an answer copied from a draft is billed once. Flash never drafts, so it never bills reasoning beyond the model's own.
Will there be open-weight models?
Our expert-transplant and weight-merging tracks are in the lab. Nothing has been released. When something is ready, it will appear here and in the changelog.
Why is Max slower than a single model?
Because it runs several drafts, a merge and an audit. Once two drafts land it waits only max(8 s, half the elapsed time) for stragglers, which bounds the delay but does not remove it.